Computational Proxemics: Simulation-based analysis of the spatial patterns of conversational groups

Abstract

In real-world conversational groups, interactants adjust their body position and orientation relative to one another in order to see and hear clearly. We use an agent-based modelling approach to compare alternative models for simulating the spatial patterns of conversational groups. The models are based on simple rules that control the movement, positioning, and orientation behaviour of individual agents, which in turn leads to the emergence of agent clusters. We identify which model alternative produces agent clusters with characteristics typical of real-world conversational groups. The centroid-based approach, where agents readjust their position and orientation with respect to the group centroid point, is a commonly used method to simulate conversational groups, but has not been empirically validated. This thesis replicates, evaluates, and validates the centroid-based model in a systematic way. Another model, where agents perform positional-orientational readjustments to see as many neighbours as possible within a 180 field of view, called the field-of-view approach is proposed, implemented, evaluated, and validated. Analysis of the spatial patterns of conversational groups has hitherto mostly relied on visual verification. We, novelly, use a combination of qualitative and quantitative methods to analyse the spatial patterns of conversational groups. Evaluations show that the field of- view model and centroid-based model produce agent clusters with significantly different social, spatial, and temporal characteristics. Validation is performed using a dataset which captures the spatial behaviour of 21 participants for the entire duration of a party. This validation shows that the characteristics of agent clusters resulting from the field-of-view model most closely reflects the characteristics of real-world conversational groups. We also show that a local neighbourhood influence works better than an extended neighbourhood influence to simulate conversational groups. The influence of objects in the environment on the spatial patterns of agent clusters are also discussed.